Comment 2 for bug 30415

Celso Providelo writes ("[Bug 30415] Re: soyuz upload system pays attention to ftp sessions"):
> Following you suggestion, I think, we need to support a single
> control file inside the upload directory which controls its
> processing, of course, it needs support from 'dput' side. It would
> be something like:

Why not process each .changes as soon as the files it lists have
arrived ? A .changes is detailed enough to tell whether the upload is
complete. Stale files can be expired, and it is safe to allow
overwrites - or you could even do something fancy like keep both
versions of the file and pick the one that matches the md5sum in the
.changes.

In fact, there is not really any reason to support directory listings
on the incoming directories, nor downloads from them. Once we realise
this, then it becomes obvious that there is no need for the _filename_
to be a unique key for finding an uploaded file.

Non-.changes files which have been uploaded should be indexed by
(filename,size,md5sum), as that's the information which is listed in
the .changes. .changes files can be indexed by the md5sum of the
.changes, and if two .changes with the same name are uploaded then
whichever one gets all of the files is processed - if both get all of
their files then both are processed, possibly causing an error if that
would violate in-archive uniqueness.

If you do this then you don't need all the palaver with directories
either, because the .changes will say where (which distro, which
personal package archive, etc.) the files should be sent.

Ian.